{"id":7596,"date":"2025-03-24T05:24:09","date_gmt":"2025-03-24T12:24:09","guid":{"rendered":"https:\/\/meta-quantum.today\/?p=7596"},"modified":"2025-03-24T05:24:09","modified_gmt":"2025-03-24T12:24:09","slug":"nvidia-new-ai-model-n1-explained-in-detail","status":"publish","type":"post","link":"https:\/\/meta-quantum.today\/?p=7596","title":{"rendered":"NVIDIA: NEW AI Model N1 Explained – in Detail"},"content":{"rendered":"\n

Introduction<\/h2>\n\n\n\n

NVIDIA has unveiled N1, a foundation model for generalist humanoid robotics. Though relatively small with 2.2 billion parameters, this model required 50,000 GPU hours of training across 1,024 H100 GPUs. What makes N1 remarkable isn’t its size but its innovative architecture\u2014it operates across six different vector spaces, enabling robots to understand visual input, process language instructions, and generate physical actions.<\/p>\n\n\n\n

Architecture and Design<\/h2>\n\n\n\n

NVIDIA’s N1 is a foundation model for generalist humanoid robotics that operates across six distinct vector spaces. Despite having only 2.2 billion parameters (relatively small by modern standards), its innovation lies in how it bridges perception, language understanding, and physical action generation for robots.<\/p>\n\n\n\n

The six vector spaces include:<\/p>\n\n\n\n

    \n
  1. Visual Embedding Space<\/strong> – Using CLIP 2 visual encoder for image processing<\/li>\n\n\n\n
  2. Linguistic Embedding Space<\/strong> – Using a small LM2 language model<\/li>\n\n\n\n
  3. Unified Mathematical Space<\/strong> – Combining vision and language via the Eagle 2 backbone<\/li>\n\n\n\n
  4. Embodiment-specific State Space<\/strong> – Representing robot’s physical state<\/li>\n\n\n\n
  5. Action Embedding Space<\/strong> – Encoding possible robot actions<\/li>\n\n\n\n
  6. LAPA (Latent Action Pseudo Annotation) Space<\/strong> – Extracting actionable information from videos<\/li>\n<\/ol>\n\n\n\n

    Dual-System Architecture<\/h2>\n\n\n\n

    N1 employs a System 1\/System 2 approach similar to human cognition:<\/p>\n\n\n\n

      \n
    • System 2<\/strong>: A slower (~10Hz) vision-language model for processing and reasoning<\/li>\n\n\n\n
    • System 1<\/strong>: A faster (~120Hz) action generation system for real-time motor control<\/li>\n<\/ul>\n\n\n\n

      Training and Data Requirements<\/h2>\n\n\n\n

      N1 required substantial computational resources:<\/p>\n\n\n\n

        \n
      • 50,000 GPU hours of training across 1,024 H100 GPUs<\/li>\n\n\n\n
      • Training data included web data, human videos, synthetic data, and some real-world robot demonstrations<\/li>\n\n\n\n
      • NVIDIA generated 780,000 simulation trajectories (equivalent to 6,500 hours of movement) in just 11 hours<\/li>\n<\/ul>\n\n\n\n

        Key Innovation: LAPA<\/h2>\n\n\n\n

        One of N1’s most significant innovations is the LAPA system, which:<\/p>\n\n\n\n

          \n
        • Uses a Vector Quantized Variational Autoencoder to learn from videos<\/li>\n\n\n\n
        • Extracts implicit motion information between video frames without explicit robot commands<\/li>\n\n\n\n
        • Converts video observations into actionable representations for robots<\/li>\n\n\n\n
        • Creates pseudo-labels for robotic actions by analyzing human movements<\/li>\n<\/ul>\n\n\n\n

          Flow Matching for Action Generation<\/h2>\n\n\n\n

          Rather than traditional diffusion models, N1 uses flow matching which:<\/p>\n\n\n\n

            \n
          • Learns a time-dependent vector field that guides noisy action sequences toward meaningful ones<\/li>\n\n\n\n
          • Requires only four iterations during inference for real-time performance<\/li>\n\n\n\n
          • Efficiently translates high-level goals into precise motor commands<\/li>\n<\/ul>\n\n\n\n

            Operational Workflow<\/h2>\n\n\n\n
              \n
            1. Robot receives sensory input and language instructions<\/li>\n\n\n\n
            2. Vision-language model interprets the environment and task<\/li>\n\n\n\n
            3. Robot’s current state is encoded<\/li>\n\n\n\n
            4. Diffusion Transformer generates appropriate actions<\/li>\n\n\n\n
            5. Actions are decoded into specific motor commands<\/li>\n\n\n\n
            6. Robot executes commands and the cycle repeats<\/li>\n<\/ol>\n\n\n\n

              Limitations and Challenges<\/h2>\n\n\n\n
                \n
              • Computationally intensive, requiring specialized hardware<\/li>\n\n\n\n
              • Heavy reliance on synthetic data may create artifacts<\/li>\n\n\n\n
              • Uses relatively low-resolution image processing (224\u00d7224 pixels)<\/li>\n\n\n\n
              • Complex body movements require sophisticated recognition and translation<\/li>\n<\/ul>\n\n\n\n

                Applications<\/h2>\n\n\n\n

                As a foundation model for humanoid robotics, N1 could potentially be used for:<\/p>\n\n\n\n

                  \n
                • General-purpose humanoid robots that can understand natural language instructions<\/li>\n\n\n\n
                • Robots that can learn new tasks by watching demonstrations<\/li>\n\n\n\n
                • Systems that bridge the gap between perception and physical action<\/li>\n\n\n\n
                • Research platforms for advancing embodied AI<\/li>\n<\/ul>\n\n\n\n

                  N1 represents NVIDIA’s commitment to open foundation models in robotics while showcasing their hardware capabilities. By combining relatively simple mathematical operations across multiple vector spaces, N1 demonstrates an approach to creating robots that can perceive, reason, and act in the physical world.<\/p>\n\n\n\n

                  Video about NIVIDIA N1:<\/h2>\n\n\n\n
                  \n